U.S. Bank
As a product management intern on U.S. Bank's AI Center of Excellence, I redesigned how business teams request AI work, ran a workshop that mapped what was worth reusing, and built the roadmap that came out of it.
The AI Center of Excellence partners with business teams across U.S. Bank to scope, build, and deploy AI. Getting that process started required teams to fill out a long intake form. Many of them aren't technical, and the form assumed familiarity with the AI development process that most requesters didn't have.
My summer started with one project: redesign the intake experience so non-technical teams could get started without knowing the right questions in advance. It ended with two: a centralized platform prototype that senior leadership approved, and a reusability roadmap that came out of a workshop I ran with four AI teams.
I ran 20+ interviews across the groups that touched the intake: AI champions, risk management, and teams that had recently submitted requests. I also reviewed the existing guidance and communications.
The pattern was consistent. This wasn't a form problem; it was a knowledge problem. Requesters were confused about which path their request should take. Answers to common questions were spread across multiple sources and hard to find. The same questions came back repeatedly, adding overhead to the AICoE team. And almost everyone wanted to know what would happen next. The form asked people to fill in the right answers before anyone had given them the right questions.
Requesters shouldn't have to know which questions to answer. The system should ask them.
The end goal was a conversational intake agent: someone describes what they need in plain language, and the system works out the rest. The question was how to get there credibly, without building on knowledge sources that hadn't been vetted yet.
The Request Guide centralizes the fragmented information and establishes what's authoritative. The FAQ chatbot tests whether people can find answers in that content through natural-language questions. The platform's intake agent then uses the same vetted sources to run a conversation instead of a form. Each step earned the right to build the next one.
the platform also has a portfolio dashboard for senior leadership and a collaborative workspace for risk and AI champion teams. one platform, three views.
The FAQ chatbot could maximize coverage (answer more, more loosely) or maximize trust (answer less, with citations, with escalation when it can't). In a regulated financial institution, a confident wrong answer is worse than no answer.
Every answer cites its source. When confidence is low or the question is out of scope, the chatbot routes to a person instead of guessing. The design questions I worked through before building: which sources are authoritative? How do we stay accurate as content changes? What happens when the answer isn't there? Who owns the knowledge sources long term? Those questions shaped the platform more than any technical choice.
What information does the AICoE need from me to process a request?
The intake process has three stages. Stage one is a business need summary — a plain-language description of what you're trying to do and who it affects. Stage two is scoping, run with an AI champion. Stage three is formal intake, which requires a completed risk and feasibility checklist.
The second project: AI teams across the AICoE had built components that other teams were rebuilding from scratch. Reusable assets existed but weren't consistently documented or discoverable. With my mentor, I designed a workshop to find out what was worth reusing and why reuse wasn't happening.
We ran it as a game: rounds, points, a live leaderboard. Four teams analyzed 44+ assets and cast 427+ votes on the most valuable ones. The competition got teams to surface blockers honestly instead of diplomatically, and the vote totals turned opinions into data. The results told a clearer story than any survey would have.
The workshop showed that reuse wasn't a search problem. 67% of feedback pointed to documentation and asset quality as the main blocker. Teams couldn't find assets they knew existed, but they also couldn't use assets they found because the documentation wasn't there.
Phase 1 is three assets, fully packaged: spec and flow diagram, canonical use case, usage patterns, inputs and outputs, limitations, named owner, support path. The goal isn't scale — it's validating the standard. If teams can find and consume the first three, the portal makes sense. If they can't, the standard needs to change first.
"documentation standard" sounds boring. what it means: someone can pick up an asset they've never seen and use it without asking anyone.
centralized platform prototype, including the intake agent, by senior leadership
votes cast across 44+ assets in the reusability workshop
presented to senior leadership; portal MVP planned for a later phase
The handbook was delivered, the chatbot was prototyped, the workshop ran, and the roadmap was built. What I took away was less about the outputs and more about where AI products actually break: not at the model, but at information, ownership, governance, and implementation. The model is the last piece of the puzzle, not the first.
I went into every conversation expecting to find answers and kept finding better questions instead.